Quantum NLP Enhances Accuracy of Bank Policy Text Classification

Bank Policy Management and Text Classification Challenges

Bank policy documents are numerous, complex, and highly inter‑related, making traditional management approaches prone to compliance gaps and inefficient classification. Conventional natural‑language text classifiers struggle to capture deep patterns in such high‑complexity data, especially under small‑sample conditions.

Quantum Machine Learning Background

Quantum Machine Learning leverages quantum superposition and entanglement to generate richer feature representations. Prior work shows quantum clustering and classification outperform classical methods in domains such as financial risk detection and image recognition. Servedio et al. [1] demonstrated the equivalence of learnability between quantum and classical information, while Kak introduced quantum neural networks in 1995 [2]. Chen et al. proposed the QCNN model in 2017 [3], laying the foundation for quantum‑enhanced NLP (QNLP) and quantum support vector machines (QSVM) for text classification.

QCNN Model for Bank Text Classification

The proposed architecture (Figure 1) embeds raw text into dense vectors, then feeds them to a QCNN that applies quantum convolution kernels in Hilbert space. The QCNN uses a 4‑qubit device: input features are encoded via Y‑rotations parameterized by w, followed by a layer of 12 random quantum operations, rotation and CNOT gates, and Pauli‑Z measurement. Text is partitioned into 2×2 blocks, each processed through the quantum encoder, quantum layer, and measurement to produce quantum feature vectors that capture deep semantic and structural patterns.

Dataset and Experimental Setup

Bank policy titles from domestic financial regulators were pre‑processed and manually labeled into six categories (e.g., “Product & Service”, “Regulatory & Risk Control”). Labels were further grouped into two high‑level classes. Experiments focused on semi‑supervised learning with only 50 or 100 samples per class. The QCNN was compared against a classical TextCNN baseline (Kim 2014) [4] using identical training hyper‑parameters: 20 epochs and batch size 4.

Results and Evaluation

Table 1 (Figure 2) shows that QCNN outperforms TextCNN on Accuracy, F1_macro, and F1_micro. With 50 samples, QCNN achieves 0.63 Accuracy (+7 pp), 0.52 F1_macro (+5 pp), and 0.77 F1_micro (+8 pp). With 100 samples, QCNN reaches 0.70 Accuracy (+6 pp), 0.71 F1_macro (+12 pp), and 0.78 F1_micro (+6 pp). These gains indicate stronger classification precision and generalization, especially in small‑sample regimes.

Conclusion

Applying QCNN to bank policy text classification demonstrates that quantum‑enhanced feature extraction can significantly improve accuracy and robustness over classical deep‑learning models. The study validates quantum natural language processing as a viable tool for reducing compliance risk and advancing digital, intelligent bank policy management.